In June 2026, a federal judge in Mississippi sanctioned four attorneys who had filed case citations their AI tools invented. One worked at a firm whose policy already required that AI work be verified. Transportation work carries the same exposure, with potentially serious consequences. A stamp and/or signature says the professional has reviewed the document and stands behind it, but it says nothing about whether AI helped write it. AI fails in two ways that ordinary review is not built to catch. It states wrong things with complete confidence: a standard that does not exist, a signal timing value from the wrong state's manual. And when asked to tidy up technical text, it can tweak one word, such as turn a "shall" into a "should", and in doing so hand back a paragraph that it thinks reads better but now means something different.
The following is a summary of this section’s findings and recommendations:
| Stage | The check |
|---|---|
| Before | Decide whether the task is a fit for AI. Fine for AI: memos, summaries, outlines. Heightened scrutiny: interpreting standards, regulatory content. Not without significant review: final engineering judgment, safety-critical calculations. |
| During | Make the AI output checkable. Put the governing standard in the prompt. Ask for sources and check them as they appear. Log which tool, what task, which sections. |
| After | Verify against the primary source, not against how confident it sounds. Do this for every citation, spec, and number. When AI has rewritten text from a standard or an earlier deliverable, read the two side by side. Closely scrutinize use of modal verbs (e.g. shall vs should), negations, numbers and units, defined terms, and qualifying clauses like "where conditions permit." |
| Sign-off | Three people, three jobs. The user does a first pass. The QC reviewer reviews the whole deliverable, not just the flagged parts. The signing professional is told where and how AI was used. |
For stamped work, consider adding additional review or keeping the findings AI-free.
Before a work product is finalized:
- Task confirmed as a fit; approved tool used.
- Standard given to the tool; sources checked as they appeared.
- AI use logged: tool, task, sections.
- Every citation, spec, and number verified against the primary source.
- AI-edited text compared to the original for meaning drift.
- Full deliverable reviewed as one document.
- First-pass review, QC review, signing professional informed.
The rest of the section covers accountability, the four QA/QC stages, added safeguards for high-stakes deliverables, and a checklist for verifying outputs.
A four-step QA/QC process
The four stages track the work itself: deciding whether AI fits the task, using it, verifying the output, and signing off. But first, an overview of your responsibilities as a professional.
The professional remains accountable
In June 2026, a federal judge in Mississippi barred two attorneys from his court for two years and fined four. Both sides of a civil lawsuit had used artificial intelligence (AI) tools, filed the case citations those tools produced, and never checked whether the cases were real. They were not. One of the four was a partner at a firm whose written policy required AI research to be verified. The other side's attorney told the judge she had not known AI could invent cases at all (Vigdor 2026).
Transportation organizations face similar risks. AI can shape what goes into a stamped design report, an approved plan, or anything else a professional signs their name to, and staff are already using these tools with or without oversight.
No matter how AI is used to produce a work product, the professional who reviews, approves, and delivers it remains responsible for its accuracy, completeness, and compliance with applicable standards. AI does not change who owns the work.
A P.E. stamp has always meant, “I have reviewed this and I stand behind it.” It certifies the document itself, not the competence of whoever, or whatever, helped produce it. An AI tool is best treated the way the profession already treats existing software for calculations and design: useful, but the professional remains responsible for the inputs and for checking that the output makes sense.
As these tools spread, professional work shifts from executing a task toward supervising a tool that executes it (Armstrong and Shah 2026). The following QA/QC framework has been devised to keep the professional accountable and avoid passive review where the human signs off on whatever the tool produced.
Stage 1: Before using AI
Not every task is a good fit, and screening the task up front is cheaper than catching errors in review. Below are three useful screening steps (GSA 2026; USDOT 2025b).
- Step 1
Check that the task is generally appropriate for AI assistance.
Those generally include: drafting and editing prose such as memos and correspondence, summarizing long documents, organizing references, and first-draft outlines or checklists. The user can see the whole output and judge it, and an error is usually obvious.
- Step 2
Use AI with heightened scrutiny.
For tasks involving: Interpreting design standards or specifications, summarizing technical findings, preparing content for regulatory submissions, and anything where the AI output will be carried into a deliverable without substantial rewriting. The output looks authoritative, but the errors are easy to miss, so verification must be deliberate.
- Step 3
Identify tasks where AI use is not recommended without significant additional review.
These are final engineering judgments, novel design situations without clear precedent, safety-critical calculations, and anything where a wrong output could directly cause harm and the error would be hard to detect. AI may assist around the edges, but it should never be the source of the judgment.
In practice, keep a short list of approved AI use cases and task types, kept current by a designated AI lead and ask the suitability question at the outset (NIST 2023).
Stage 2: During AI use
Prompting for verifiable outputs
Be specific. Vague requests produce vague outputs that are hard to check, so put the relevant standard, specification, or requirement in the request itself, which gives something concrete to verify the answer against. Ask the tool to cite its sources, and keep in mind that these tools will still hallucinate sources even when prompted, and can make mistakes even when asked to comply with a relevant standard. Manually verify every citation and treat the output as a starting point that needs independent verification, not a near-final draft that needs light editing.
Vetted tools
Use only tools your organization has vetted. That said, many organizations have not yet rolled out vetted AI tools applicable to tasks that staff are seeking to apply AI to. In a 2025 survey by MIT, workers at more than 90% of companies reported using personal AI tools while only 40% of those companies had bought an official subscription; where the approved option does not fit the task, staff route around it (MIT NANDA 2025). When this happens, licensed standards or client-confidential material can end up in a tool that retains or trains on it. Organizations should give staff vetted enterprise access to the range of tools their work actually requires (see Governance and Secure Use sections for further guidance in this area).
Documentation
Keep a record of AI use for any professional deliverable: namely which tool was used, the general nature of the task, and which portions were AI-assisted. For most work a brief note in the project file is enough although high-stakes deliverables need more. This record tells the reviewer that AI was used, so the Stage 3 checks get applied, and it leaves a trail if questions surface later (NIST 2024a).
Source tracing
Trace sources during use. If the AI references a specific standard, specification, code section, study, or case, note and check the reference the moment it appears rather than chasing it down during review.
Stage 3: After AI output
This stage guards against two failures: AI getting something wrong, and AI changing what something says in a misguided attempt to make it more readable. Each failure mode is covered in turn, below.
Failure mode 1: Factual errors and fabricated references
AI tools state wrong things with complete confidence. An AI-assisted report might cite a standard that does not exist, or cites a real one and misquotes its values. An AI-assisted signal-timing requirement might come from the wrong state's manual. An AI-assisted design speed might come from the wrong design class. These errors can be hard to spot because none of them look any different from correct text around them.
To guard against this, you must verify every citation, standard reference, specification, and numerical value produced by AI against the primary source. This is the same level of discipline that has always been the professional standard: confirm the result against an objective reference rather than against how confident it sounds (Grace et al. 2026).
Failure mode 2: Meaning drift
AI can try to make text read more smoothly in a way that does not reliably preserve precise technical or legal meaning, especially when the original language was chosen with care. In engineering, a single word can completely change the meaning and obligation. "Shall," "should," and "may" separate what is required, recommended, and merely permitted. The MUTCD is built on that distinction, so a single substituted verb can turn a requirement into a suggestion.
When AI has edited or rewritten text that came from a technical document, a standard, or a previously reviewed deliverable, compare the AI version against the original instead of reading it on its own. Meaning drift is nearly invisible in isolation because the rewritten text sounds correct even when it has changed. Pay particular care to these categories:
- Modal verbs: shall, should, may, etc.
- Negations: not, never, except.
- Numerical values and units.
- Defined terms, which AI may swap for synonyms that carry a different technical meaning.
- Qualifying clauses such as “in most cases,” “subject to engineering judgment,” or “where conditions permit,” which AI tends to drop for cleaner prose.
The review itself
Apply the same treatment to AI-assisted content as you would to plan review, i.e. the reviewer checks the submittal against the design intent, verifies the references, and confirms nothing was lost or changed in translation. The project-file note from Stage 2 tells the reviewer where AI was used, but it should not shrink the review to a spot-check of the flagged passages; the reviewer must approach the full deliverable as one document that needs checking.
Stage 4: Sign-off and accountability
The structure for the sign-off stage has three tiers that map onto existing roles within engineering organizations (NIST 2023).
Tier 1The AI user
The person who used the tool owns the Stage 2 practices and owes a thorough first-pass review before the output goes anywhere. This is the person closest to how the output was produced, and is the “cheapest” place to catch an error.
Tier 2The QC reviewer
The QC reviewer examines the full deliverable as a work product, using the same peer review the organization applies to anything else, plus the Stage 3 checks on AI-assisted sections (whose locations were noted in Stage 2). The reviewer checks the document itself, and AI does not remove the need for this step.
Tier 3The signing professional
The P.E. or equivalent who signs bears ultimate accountability. That professional should be told how and where AI was used and have access to the Stage 2 documentation.
At the organizational level
Designate an individual to gather report-outs on what errors AI tools are making, to keep the organization informed of what review needs to be made. This individual would track incidents where AI-assisted work produced an error, update guidance as tools change, and coordinate with whomever maintains the approved-tool list (Microsoft 2025).
High-stakes deliverables
Some deliverables need additional safeguards: those where an error could affect public safety, carry regulatory or legal standing, create financial or liability exposure, or involve the application of a professional seal.
For example, in engineering: safety-critical design elements such as geometric design and sight distance; traffic impact studies submitted to agencies for approval; and construction plans and specifications stamped by a P.E.
For these deliverables, consider these additional safeguards:
- Document every AI-assisted section explicitly, not with a general note.
- Require a second independent review of AI-assisted content, separate from standard QC and done by someone who did not produce the document.
- Require the reviewing professional to sign off that AI-assisted sections were verified against primary sources.
- Consider whether any section should be AI-free. This is often most appropriate for passages that could be quoted out of context, such as executive summaries and findings statements.
Federal agencies govern AI in proportion to risk in the same way: the highest-impact uses get additional testing, human review, and continuous monitoring, and they keep a non-AI fallback process available (GSA 2026; USDOT 2025c).
Output verification checklist
The above guidance has been summarized below in a handy checklist.
Before starting (Stage 1)
During use (Stage 2)
Verifying the output (Stage 3)
Sign-off (Stage 4)
Sources cited on this page are listed in full on the References page.